GHNN: Graph Harmonic Neural Networks for semi-supervised graph-level classification.

Area of Science:

• Machine Learning
• Graph Representation Learning

Background:

• Graph classification predicts properties of entire graphs, a key task in graph learning.
• Existing methods like graph convolutional networks (GCNs) implicitly learn topology, while graph kernels explicitly use structural knowledge.
• Label scarcity in real-world data necessitates semi-supervised approaches.

Purpose of the Study:

• To propose a novel semi-supervised learning framework for graph classification that addresses label scarcity.
• To combine the strengths of graph convolutional networks and graph kernels for enhanced graph topology exploration.
• To leverage unlabeled data effectively through novel loss functions.

Main Methods:

• Introduced Graph Harmonic Neural Network (GHNN), integrating GCN and graph kernel network (GKN) modules.
• Developed a harmonic contrastive loss and a harmonic consistency loss to harmonize module training.
• Prioritized high-quality unlabeled data to reconcile predictions between GCN and GKN modules.

Main Results:

• GHNN demonstrated superior performance on various benchmark datasets compared to existing methods.
• The integrated approach effectively leveraged both labeled and unlabeled data for improved graph classification.
• Mutual enhancement between GCN and GKN modules led to comprehensive graph topology exploration.

Conclusions:

• GHNN successfully overcomes label scarcity in semi-supervised graph classification.
• The proposed harmonic losses effectively utilize unlabeled data, improving model robustness and accuracy.
• Combining implicit and explicit graph topology learning offers a powerful strategy for graph classification tasks.